Crime-detection automatic speaker verification and identification (CASVI) system

Abstract

The crime-detection automatic speaker verification and identification (CASVI) system, using a new spectral and cepstral analysis technique, is described. Robust speaker recognition algorithms, the operation modes, the control windows, the testing methodology, and the testing results of the system are presented. Research of the system has shown it can efficiently work with strongly corrupted utterances, while the corruption of the utterances being compared may be considerably different. For any utterances, permissible frequency distortions are 35 dB in the operating band and the permissible signal-to-noise ratio of additive noise is 12 dB. In these conditions, the CASVI system using 12-s speech utterances demonstrates the possibility to verify and identify persons on the level of 90% of correct acceptance when the false-alarm rate is 10%. The verification and identification time by the system takes some minutes. The CASVI system does not require high qualification of an expert. The system has a very simple interface that allows working in full automatic mode and forming reports without any special schooling. The CASVI system is a software product installed on a Pentium-based PC running with Windows 95.